The most-wanted AI model of the summer just came back online — and it may already be the second-best model Anthropic has, behind one the public has never seen.
That’s the strange shape of this week. Claude Fable 5 is being restored today after an 18-day government blackout. OpenAI’s GPT-5.6 is queued right behind it, waiting on Washington’s green light. And in the middle of all this, a credible rumor says the truly frontier model — more capable than any of them — is already trained and sitting idle on Anthropic’s servers.
If you only take one thing from this: the AI you’re allowed to use is now a curated slice of the AI that actually exists. Here’s the full picture, in three parts.
Fable 5 is back. GPT-5.6 is next. And Anthropic reportedly already has something stronger.
The most-wanted model of the summer is online again — and it may already be the second-best model Anthropic has, behind one the public has never seen. The AI you’re allowed to use is now a curated slice of the AI that exists.
Restored on Claude platform, Claude.ai & Code. Up to 50% of weekly limits through July 7. Was briefly the benchmark king — now returns with new safeguards & possible ID checks.
Previewed June 26 to only ~20 government-vetted partners; general release “in coming weeks,” pending Washington’s nod. Cheaper than Fable — roughly half the price.
OpenAI · compute-heavy
OpenAI · flagship
the tie — “Fable-5 level”
Anthropic · GA fallback
On June 21, ~9 days into the blackout, AI analyst Andrew Curran said on X that Anthropic had already finished training a more capable Mythos successor — possibly shipping as Mythos 5.1 / 6, possibly staying internal. Anthropic hasn’t confirmed it. But it’s not baseless: an unreleased Mythos Preview already sits above the public tier — OpenAI even benchmarks Sol against it. The pattern is real even if the specific model isn’t proven.
Stack it up and the shape is clear: what the public can use — Fable 5 today, GPT-5.6 in weeks, whatever clears the gate next — is a permissioned, curated slice of what these labs have actually built. A stronger tier is almost always one step ahead, behind a government gate or a lab’s caution — and both companies are pushing to make that review process permanent. For builders the instruction is blunt: don’t chase “the best model.” Build so you can swap whichever one you’re allowed to use this week — because that list keeps changing.
1. Fable 5 is back — here’s what you actually get
After the Commerce Department lifted its export controls on the evening of June 30, Anthropic began restoring Claude Fable 5 to global users on the Claude platform, Claude.ai, and Claude Code. To smooth the return, it’s being included for up to 50% of weekly usage limits through July 7 on Pro, Max, Team, and selected enterprise plans, with re-enablement on AWS, Google Cloud, and Microsoft Foundry to follow.
Two caveats worth flagging. First, it’s not yet fully clear whether subscribers get back the free run of Fable they were briefly promised, or whether it returns partly behind usage credits and identity checks — Anthropic has signaled tighter access controls as part of the deal. Second, the model returns with new strings attached: Anthropic agreed to proactively detect and address security risks, work with the government on release protocols, and report malicious activity, and it added a safeguard that blocks the specific jailbreak officials worried about roughly 93% of the time. Its high-capability sibling Mythos 5 — the same underlying model with fewer guardrails — has been restored to a limited set of US organizations (reportedly 100-plus institutions defending critical infrastructure) through the Glasswing program, with access meant to widen over time.
Why the fuss over one model? Because in its three days of public life, Fable 5 was arguably the strongest thing anyone could buy. It topped coding leaderboards; in Anthropic’s own early testing, Stripe reportedly used it to overhaul a 50-million-line codebase in a single day; one widely read newsletter called it the best coding model in the world before it vanished. Losing it hurt, which is exactly why its return is news.
AIPI-Lite Smart AI Companion, Customizable AI Character with Voice Cloning, GPT Model Access, Memory System, Real-Time Reactions, Subscription Required
Risk-free Trial Requires Subscription: After purchase, please follow our instructions to subscribe for membership, activate features, and unlobck…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
2. GPT-5.6 is next in line — pending Washington’s nod
Here’s the part that reframes Fable’s comeback as a race rather than a victory lap. On June 26, OpenAI previewed GPT-5.6 — a three-tier family (Sol the flagship, Terra the mid tier, Luna the efficient tier, plus a compute-heavy Sol Ultra mode). But it launched the same way Mythos returned: as a limited preview for roughly 20 government-vetted partners, at Washington’s request, with general availability promised “in the coming weeks” — and OpenAI publicly said it doesn’t think this kind of gated process should become the norm.
Now, the “GPT-5.6 is on Fable-5 level” claim — because the precise version matters. On OpenAI’s own headline benchmark, Terminal-Bench 2.1, the numbers land like this: GPT-5.6 Terra ties Claude Fable 5 at about 84%, while the Sol flagship pulls ahead at 88.8%, and Sol Ultra stretches to 91.9% (for reference, Claude Opus 4.8 sits at 78.9%). So the honest read is: GPT-5.6’s mid tier matches Fable 5 on this test, and its top tier beats it — at roughly half the price ($5/$30 for Sol versus Fable 5’s $10/$50). The large asterisk: these are vendor preview figures on a benchmark OpenAI chose, narrow by design, not yet independently verified, with the fuller evaluation suite deferred to general availability. Treat them as an upper bound until third parties confirm. OpenAI led with agentic and cyber numbers and held the rest back — and cyber capability is precisely what triggers the new government gate, which is why GPT-5.6 became the first marquee model to pass through it.
AI Engineering: Building Applications with Foundation Models
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
3. The twist: a stronger model may already exist
Now the rumor — and it should be read as exactly that, a rumor, attributed and unconfirmed.
On June 21, roughly nine days into the blackout, AI analyst Andrew Curran said on X that Anthropic had already finished training a more capable successor to Mythos 5 — a model that could ship as something like Mythos 5.1 or Mythos 6, or simply stay internal. His broader point was blunt: pulling models from public use does essentially nothing to slow frontier development, and by freeing up resources may even nudge it faster. Several outlets picked the claim up; Anthropic has not confirmed it, and no name, benchmark, or release date is established. Take it with appropriate salt.
But it isn’t baseless, and here’s the corroborating detail most coverage skips: there is already a known, unreleased Mythos Preview model — the original April Glasswing release — that sits above the public tier, and OpenAI’s own GPT-5.6 materials benchmark Sol as competitive with that unreleased Mythos Preview on an exploit test. In other words, the pattern is confirmed even if the specific “next model” isn’t: the most capable systems these labs build routinely stay behind the public line, released late, partially, or not at all. A stronger Anthropic model existing right now would be consistent with everything on the record — which is why the rumor is plausible, not that it’s proven.
AI Programming Made Practical: A Step-by-Step Guide to Building AI-Powered Applications, Writing Better Code Faster, and Using Modern AI Tools with Confidence
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What it all means: the frontier is now curated
Stack the three parts together and the shape is unmistakable. What the public can actually use — Fable 5 today, GPT-5.6 in a few weeks, whatever passes the gate next — is a permissioned, curated slice of what these companies have built. Models get switched off (Fable), staggered to vetted partners (GPT-5.6), or held back entirely (Mythos Preview, and possibly its successor). A more capable tier is almost always one step ahead of you, behind a government gate or a lab’s own caution.
That’s now structural, not incidental. Both Anthropic and OpenAI are actively pushing Washington to codify a model-review process — building on a June 2 executive order that set up voluntary government vetting of frontier models — which means the improvised gating of the last month is on track to become permanent infrastructure. The competitive race is real (GPT-5.6’s pricing and Sol’s benchmark lead are genuine pressure on Anthropic), but it now runs through a national-security checkpoint that decides who gets what, when.
For anyone building on these systems, the takeaway is the same one this month keeps teaching: the “best model” is a moving, permissioned target. Marrying your product to a single model or a single lab’s release calendar is a bet on a variable you don’t control. Portability — multiple providers, tested fallbacks, and open-weight or self-hosted options where they fit — is no longer a nice-to-have; it’s the only posture that survives a frontier you can’t fully access.
Evals for AI Engineers: Systematically Measuring and Improving AI Applications
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The take
Fable 5’s return is real and worth celebrating if you rely on it. GPT-5.6 is coming, and on the numbers it’s a genuine rival — cheaper, and at the top end faster on the benchmark OpenAI chose to show. And somewhere behind both, if the rumor holds, sits a model stronger than anything you can call an API against today.
The uncomfortable truth underneath the excitement is that the public no longer sees the actual frontier — it sees the part of the frontier that’s been cleared for release. Both readings of that are defensible: gating the most dangerous capabilities before they ship has a real safety logic, and so does the worry that an opaque, permissioned frontier keeps the best tools from the people who could use them well. What isn’t debatable is the instruction for everyone downstream. Don’t chase “the best model.” Build so you can swap whichever one you’re allowed to use this week — because that list is going to keep changing.
Sources: Anthropic and Commerce Secretary Howard Lutnick (statements via X); CNBC, Axios, Semafor, Forbes, Al Jazeera, 9to5Mac (Fable 5 restoration, terms, Mythos 5 limited release); OpenAI GPT-5.6 preview coverage via DataCamp, Lushbinary, BenchLM, and explainx (tiers, Terminal-Bench 2.1 figures, pricing, government-gated preview); Andrew Curran via X as reported by SaaSCity and Yellow.com (the unconfirmed rumor of a trained Mythos successor); TechCrunch and Anthropic (Mythos Preview, tier structure). Benchmark figures are vendor-reported preview numbers unless otherwise noted and are not independently verified; the successor model is an unconfirmed rumor. Details reflect reporting as of July 1, 2026 and are developing. Analysis and opinions are the author’s and not investment advice.
